Method and system for multi-source algae image target detection

Patent metadata

Field	Value
Publication number	US-12412389-B2
Application number	US-202218063781-A
Country	US
Kind code	B2
Filing date	Dec 9, 2022
Priority date	Sep 2, 2022
Publication date	Sep 9, 2025
Grant date	Sep 9, 2025

Abstract

The present disclosure relates to a method and system for multi-source algae image target detection, and relates to the field of monitoring of algal bloom events in fresh water. The method includes first crawling images of algae of a selected species by using a built automated algae crawling tool, where the images include all formats; classifying and labeling algae in the algae images, and forming a source domain dataset by using all the classified and labeled algae images; performing transfer learning by using a faster recurrent revolutional neural network (Faster RCNN) with reference to a target domain dataset, to obtain a multi-source algae image target detection model; and finally performing identification and classification by using the multi-source algae image target detection model.

First claim

What is claimed is:

1. A method for multi-source algae image target detection, comprising:
building an automated algae crawling tool;
crawling multi-source algae images by the automated algae crawling tool based on algae name index;
wherein after the crawling multi-source algae images by using the automated algae crawling tool based on an algae name index, the method further comprises:
inputting all the algae images into a pre-trained binary classifier to determine whether the algae image is algae, and obtaining a first output result;
deleting the algae image if the first output result indicates that the image is not alga; or
retaining, if the first output result indicates that the image is algae, the clarity greater than the clarity threshold and a proportion of algae in the image greater than the threshold;
determining, by using a target detection algorithm, whether a foreign substance exists in the retained algae image, and obtaining a second output result;
retaining the algae image if the second output result indicates that the image is not alga; or
deleting, if the second output result indicates that the image is algae, algae images whose foreign substance is determined, by using an ablation experiment, as interfering with subsequent network learning; and
selecting a preset quantity of algae images from the retained algae images, and performing data enhancement on the retained algae images through cropping, rotation, and scaling:

framing a bounding box of algae in an algae image by YOLO v3 target detection algorithm, and assigning a label of a crawler index, to form a source domain dataset;
obtaining a target domain dataset composed of unlabeled algal images;
performing transfer learning by using a Faster Recurrent Convolutional Neural Network based on the source domain dataset and the target domain dataset, to obtain a multi-source algae image target detection model;
inputting a to-be-detected algae image into the multi-source algae image target detection model, and outputting a pixel-level classification result.

2. The method for multi-source algae image target detection according to claim 1 , wherein the automated algae crawling tool comprises a uniform resource locator (URL) manager, a scheduler, a web page downloader, and a web page parser;
the scheduler is used to obtain a to-be-crawled URL address from the URL manager, and transmit the to-be-crawled URL address to the web page downloader;
the web page downloader is used to download a web page based on the to-be-crawled URL address, convert the web page into a web page string, and then transmit the web page string to the web page parser by using the scheduler;
the web page parser is used to parse the web page string, obtain download addresses of algae images by regular expression, and crawl algae images from open-source Chinese and English websites in a semi-automatic manner based on the download addresses; and
the scheduler is further used to receive the crawled algae images.

3. The method for multi-source algae image target detection according to claim 1 , wherein the performing transfer learning by using a Faster RCNN specifically comprises:
freezing the initial setting convolutional layers and pooling layers of a Backbone module in the Faster RCNN, and extracting a feature map by using the Backbone module;
adjusting network parameters based on a difference between the source domain dataset and the target domain dataset; and
using a multi-kernel maximum mean discrepancy (MK-MMD) in a deep adaptation network to adapt to multiple fully-connected network layers of an RCNN network module in the Faster RCNN, to reduce a difference between a source domain and a target domain by using the MK-MMD.

4. The method for multi-source algae image target detection according to claim 1 , wherein the to-be-detected algae image is an algae image in which algae are observed under an optical microscope.

Assignees

• Univ Macau Sci & Tech

Inventors

• Li Jianqing
• Yuan Anran
• Wang Binbin
• Zou Hailin
• Wang Jierui

Classifications

• G06V2201/07

  Target detection · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/82

  using neural networks · CPC title

• G06F16/51

  Indexing; Data structures therefor; Storage structures · CPC title